This is the FHI quarterly update for January to March 2020. It was an exceptional quarter due to the challenges and restrictions created by the COVID-19 pandemic.
We implemented enhanced precautions a little while before the rest of the University and the UK as a whole kicked into action. We’re currently operating entirely online, with everybody working from home.
Toby Ord published his new book The Precipice – Existential Risk and the Future of Humanity. The book explores the science behind the risks we face and puts them in the context of the greater story of humanity: showing how ending these risks is among the most pressing moral issues of our time. And it points the way forward, to the actions and strategies we can take today to safeguard humanity’s future.
Toby also wrote an article for The Guardian newspaper about Why we need worst-case thinking to prevent pandemics.
Anders Sandberg continued work towards his Grand Futures book, presenting a series of seminars that will be available on the FHI YouTube channel in the coming weeks.
Centre for the Governance of AI
Papers published during the quarter:
The Windfall Clause: Distributing the Benefits of AI for the Common Good, by Cullen O’Keefe, Peter Cihon, Ben Garfinkel, Carrick Flynn, Jade Leung, and Allan Dafoe. By voluntarily adopting the Windfall Clause, firms would bindingly agree to donate a meaningful portion of their profits if they earn a historically unprecedented economic windfall from the development of advanced AI.
Five papers were presented at the AAAI / ACM AI Ethics and Society Conference in New York in February:
- The Offense-Defense Balance of Scientific Knowledge: Does Publishing AI Research Reduce Misuse of the Technology? by Toby Shevlane and Allan Dafoe.
- The Windfall Clause: Distributing the Benefits of AI for the Common Good by Cullen O’Keefe, Peter Cihon, Carrick Flynn, Ben Garfinkel, Jade Leung and Allan Dafoe.
- U.S. Public Opinion on the Governance of Artificial Intelligence by Baobao Zhang and Allan Dafoe.
- Near term versus long term AI risk framings by Carina Prunkl and Jess Whittlestone (CSER/CFI).
- Social and Governance Implications of Improved Data Efficiency by Aaron Tucker, Markus Anderljung, and Allan Dafoe.
Technical AI Safety
- Asymptotically Unambitious Artificial General Intelligence by Michael K. Cohen, Badri Vellambi (University of Cincinnati) and Marcus Hutter (Australian National University) was presented at AAAI.
- The Incentives that Shape Behaviour by Ryan Carey, Eric Langlois (University of Toronto), Tom Everitt (DeepMind) and Shane Legg (DeepMind) was presented at SafeAI@AAAI.
- (When) is Truth-telling Favored in AI Debate by Vojtech Kovarik (Czech Technical University) and Ryan Carey was presented at SafeAI@AAAI.
- Stuart Armstrong has characterised some issues with proposed impact measures Subagents and Impact Measures.
Biosecurity has continued its work on broader strategy alongside more direct policy work, both on biological risks in general and COVID-19 in particular.
Researchers from the team were closely involved in COVID-19 related work, such as establishing the COVID-19 diagnostic testing in the UK community study that will conduct testing in Oxford initially, to analyse and report effectiveness.
FHI staff has been asked to give advice at the highest level of government in the UK and the Czech Republic.
Research Scholars Programme
Jan participated in the AI Safety Research Programme and is advising on the governmental and industry response to Covid-19 in Czechia. He and a group of other RSPers and researchers have also launched EpidemicForecasting.org.
Carina presented at AAAI/AIES, spoke to the Mexican Senate on AI Ethics and advised on Mexico’s national AI strategy.
Cassidy has been doing work with the UK Government, and her COVID-19 community diagnostic study is launching this week.